System to Determine Compact Representation of Data

1. A method comprising: accessing training input data comprising image data; training an encoder module to produce a compact representation of the image data using the training input data and a loss function comprising: a bitrate loss, a distortion loss, and an embedding loss; and storing weight data associated with training the encoder module.

2. The method of claim 1, wherein the bitrate loss (Lr) is determined as: Lr=BPP(z)+BPP(h) where BPP(z) is bits per pixel of first output from an entropy encoding encoder, and BPP(h) is bits per pixel of second output from a hyper encoder.

3. The method of claim 1, wherein the distortion loss (Ld) is determined as one of: Ld=MSE({circumflex over (x)},x), or Ld=MSSSIM({circumflex over (x)},x) where x is a first image in the training input data, {circumflex over (x)} is a second image that is reconstructed from output of the encoder module, MSE is a mean-square error function, and MS_SSIM is a multi-scale structural similarity index measure function.

4. The method of claim 1, wherein the embedding loss (Le) is determined as: Le=cosinedistance(ê,e) where e is first embedding data based on a first image in the training input data and ê is second embedding data based on a second image that is reconstructed from output of the encoder module.

5. The method of claim 1, wherein during training the encoder module is in communication with: a quantization module, a hyper encoder module, an entropy encoding encoder module, a hyper decoder module, an entropy encoding decoder module, a decoder module, and an embedding module.

6. The method of claim 1, wherein during training the encoder module is in communication with a decoder module having second weight data; and further comprising: deleting the second weight data.

7. The method of claim 1, further comprising: accessing a first input image; determining first data based on processing the first input image using the encoder module and the weight data; and storing the first data.

8. The method of claim 1, further comprising: accessing a first input image of at least a portion of a first user; determining, based on processing the first input image using the encoder module and the weight data, first data; determining identification data associated with the first user; and associating the identification data with the first data.

9. The method of claim 1, further comprising: determining second training input data comprising a plurality of input image data; determining, based on processing the plurality of input image data using the encoder module and the weight data, a first set of first data; and training a first embedding model using the first set of first data.

10. A system comprising: a memory, storing first computer-executable instructions; and a hardware processor to execute the first computer-executable instructions to: access training input data comprising input data; train an encoder module to produce a compact representation of the input data using the training input data and a loss function comprising: a distortion loss, and an embedding loss, wherein during training the encoder module, the encoder module is in communication with one or more of: a quantization module, a hyper encoder module, an entropy encoding encoder module, a hyper decoder module, an entropy encoding decoder module, a decoder module, or an embedding module; and store weight data associated with training the encoder module.

11. The system of claim 10, wherein: the distortion loss (Ld) is determined as one of: Ld=MSE({circumflex over (x)},x), or Ld=MSSSIM({circumflex over (x)},x) where x is a first input in the training input data, {circumflex over (x)} is a second input that is reconstructed from output of the encoder module, MSE is a mean-square error function, and MS_SSIM is a multi-scale structural similarity index measure function; and wherein the embedding loss (Le) is determined as: Le=cosinedistance(ê,e) where e is first embedding data based on the first input and ê is second embedding data based on the second input.

12. The system of claim 10, the loss function further comprising a bitrate loss (L_r) that is determined as: Lr=BPP(z)+BPP(h) where BPP(z) is bits per pixel of first output from an entropy encoding encoder, and BPP(h) is bits per pixel of second output from a hyper encoder.

13. The system of claim 10, further comprising instructions to: access first input data; determine first data based on processing the first input data using the encoder module and the weight data; and store the first data.

14. The system of claim 10, further comprising instructions to: determine second training input data comprising a plurality of input data; determine, based on processing the plurality of input data using the encoder module and the weight data, a first set of first data; train a first embedding model using the first set of first data; and store second weight data associated with training the first embedding model.

15. A system comprising: a memory, storing first computer-executable instructions; and a hardware processor to execute the first computer-executable instructions to: access first training input data comprising a first image; train an encoder module using the first training input data and one or more loss functions; determine weight data associated with training the encoder module; determine second training input data comprising a plurality of input data; determine a first set of data based on processing the plurality of input data using the encoder module and the weight data; and train a first embedding model using the first set of data.

16. The system of claim 15, wherein the one or more loss functions include a distortion loss; wherein the distortion loss (Ld) is determined as one of: Ld=MSE({circumflex over (x)},x), or Ld=MSSSIM({circumflex over (x)},x) where x is a first input in the first training input data, {circumflex over (x)} is a second input that is reconstructed from output of the encoder module, MSE is a mean-square error function, and MS_SSIM is a multi-scale structural similarity index measure function.

17. The system of claim 15, wherein the one or more loss functions include an embedding loss; wherein the embedding loss (Le) is determined as: Le=cosinedistance(ê,e) where e is first embedding data based on a first image in the first training input data and ê is second embedding data based on a second image that is reconstructed from output of the encoder module.

18. The system of claim 15, wherein the one or more loss functions include a bitrate loss; wherein the bitrate loss (L_r) is determined as: Lr=BPP(z)+BPP(h) where BPP(z) is bits per pixel of first output from an entropy encoding encoder, and BPP(h) is bits per pixel of second output from a hyper encoder.

19. The system of claim 15, further comprising instructions to: access a first input image; determine first data based on processing the first input image using the encoder module and the weight data; and store the first data.

20. The system of claim 15, further comprising instructions to: access a first input image of at least a portion of a first user; determine first data based on processing the first input image using the encoder module and the weight data; determine identification data associated with the first user; and associate the identification data with the first data.